Bug 205627 – REGRESSION (r253926): fast/canvas/webgl/array-bounds-clamping.html fails

• Status: RESOLVED
• Resolution: FIXED
• Component: WebGL

[Alexey Proskuryakov]

fast/canvas/webgl/array-bounds-clamping.html started to fail after switching to ANGLE in r253926

It got marked as expected failure in this revision for WebKit1, but not for WebKit2, so it now shows up as a failure on bots.

https://results.webkit.org/?suite=layout-tests&test=fast%2Fcanvas%2Fwebgl%2Farray-bounds-clamping.html

The failure is hardware specific - the test still passes on many Mac models. I didn't check very carefully, seems to be passing on AMD and failing on Intel or Nvidia.

[Radar WebKit Bug Importer]

<rdar://problem/58229090>

[Kenneth Russell]

Possible to temporarily mark the expected failure so that r253926 isn't rolled back? Might be an option that needs to be passed to ANGLE's internal instance of its shader compiler.

[Dean Jackson]

This was a tricky test to get right - I think it might be incorrect.

Our non-ANGLE code passed SH_CLAMP_INDIRECT_ARRAY_BOUNDS into the shader configuration (see Extensions3DOpenGLCommon::getTranslatedShaderSourceANGLE)

I don't see a way to do this from the EGL API, but maybe it isn't necessary? Does Chrome use this option on Apple hardware?

[Kenneth Russell]

It seems likely that ANGLE's use of this workaround needs to be expanded. Currently it looks like it's only enabled on Android and AMD hardware. May need to be done always on Mac.

https://source.chromium.org/chromium/chromium/src/+/master:third_party/angle/src/libANGLE/renderer/gl/renderergl_utils.cpp;l=1596;drc=34fdc18ec806aa05ed2029d723c31de0d141a742;bpv=1;bpt=1

[James Darpinian]

The workaround was enabled upstream, so this will be fixed in the next ANGLE roll.

[James Darpinian]

ANGLE was updated in r255782